Tag Archives: ReRAM

Mott Memristor Chaos could make Efficient AI

Congratulations to Suhas Kumar, John Paul Strachan, and R. Stanley Williams of Hewlett Packard Labs in Palo Alto for showing not just how to make a Mott memristor, but that you can create controlled chaos with one. “We showed that this type of memristor can generate chaotic and nonchaotic signals,” says Williams, who invented the memristor based on theory by Leon Chua. An analysis of the material science and engineering of titanium sub-oxides as practiced by Williams at HPL for the production of standard memristors can be found in one of my old blog posts (http://www.betasights.net/wordpress/?p=1006).

Cross-section TEM of a Mott memristor composed of 8nm niobium dioxide layer between top layer of titanium nitride and bottom pillar of titanium nitride. (Original Image: Suhas Kumar/Hewlett Packard Labs, color commentary by Ed Korczynski)

Cross-section TEM of a Mott memristor composed of 8nm niobium dioxide layer between top layer of titanium nitride and bottom pillar of titanium nitride. (Original Image: Suhas Kumar/Hewlett Packard Labs, color commentary by Ed Korczynski)

The Figure shows a cross-section of a single Mott memristors formed by the region of the 8nm thin niobium dioxide (NbO2) layer that is between the 70nm diameter titanium-nitride (TiN) pillar functioning as bottom electrode and the blanket TiN layer functioning as top electrode.

Such a device exhibits both current-controlled and temperature-controlled (https://en.wikipedia.org/wiki/Mott_transition) negative differential resistance, and the proper choice of current and temperature can result in what I like to term “repeatable” chaos. It is repeatable in that a state can be controlably placed into or out-of chaos using non-linearities in electrical current-flow and temperature. From the abstract of the original article in Nature:

We incorporate these memristors into a relaxation oscillator and observe a tunable range of periodic and chaotic self-oscillations. We show that the nonlinear current transport coupled with thermal fluctuations at the nanoscale generates chaotic oscillations. Such memristors could be useful in certain types of neural-inspired computation by introducing a pseudo-random signal that prevents global synchronization and could also assist in finding a global minimum during a constrained search.

In a simulated circuit, an array of Mott memristors can be integrated with standard memristors to form a simulated Hopfield network (https://en.wikipedia.org/wiki/Hopfield_network). Hopfield nets seem to be some of the most apt models for human memory, so if we can just wire together a sufficient number of NbO Mott memristors with TiO standard memristors then we might be a step closer to functional AI.

Read the fine coverage at IEEE Spectrum:  https://spectrum.ieee.org/nanoclast/semiconductors/devices/memristordriven-analog-compute-engine-would-use-chaos-to-compute-efficiently

Or the Nature article behind paywall:  https://www.nature.com/nature/journal/v548/n7667/full/nature23307.html

—E.K.

PCM + ReRAM = OUM as XPoint

The good people at TECHINSIGHTS have reverse-engineered an Intel “Optane” SSD to cross-section the XPoint cells within (http://www.eetimes.com/author.asp?section_id=36&doc_id=1331865&), so we have confirmation that the devices use chalcogenide glasses for both the switching layer and the selector diode. That the latter is labeled “OTS” (for Ovonic Threshold Switch) explains the confusion over the last year as to whether this device is a Phase-Change Memory (PCM) or Resistive Random Access Memory (ReRAM)…it seems to be the special variant of ReRAM using PCM material that has been branded Ovonic Unified Memory or “OUM” (https://www.researchgate.net/publication/260107322_Programming_Speed_in_Ovonic_Unified_Memory).

As a reminder, cross-bar ReRAM devices function by voltage-driven pulses creating resistance changes in some material. The cross-bars allow for reading and writing all the bits in a word-string in a manner similar to Flash arrays.

In complete contrast, Phase Change Memory (PCM) cells—as per the name—rely upon the change between crystalline and amorphous material phases to alter resistance. The standard way to change phases is with thermal energy from an integrated set of heater elements. The standard PCM architecture also requires one transistor for each memory cell in a manner similar to DRAM arrays.

Then we have the OUM variant of PCM as previously branded by Energy Conversion Devices (ECD) and affiliated shell-campanies founded by tap-dancer-extraordinaire Stanford Ovshinsky (https://en.wikipedia.org/wiki/Stanford_R._Ovshinsky). So-called “Ovonic” PCM cells see phase-changes driven by voltage pulses without separate heater elements, such that from a circuit architecture perspective they are cross-bar ReRAMs.

Ovshinsky et al. successfully sold this technology to industry many times. In 2000, it was licensed to STMicroelectronics. Also in 2000, it was used to launch Ovonyx with Intel investment (http://www.eetimes.com/document.asp?doc_id=1176621), at which time Intel said the technology would take a long time to commercialize. In 2005 Intel re-invested (http://www.businesswire.com/news/home/20051019005145/en/Ovonyx-Receives-Additional-Investment-Intel-Capital). Finally in 2009, Intel and Numonyx showed a functional 64Mb XPoint test chip at IEDM (http://www.eetimes.com/document.asp?doc_id=1176621).

In 2007, Ovonxyx licensed it to Hynix (http://www.eetimes.com/document.asp?doc_id=1167173), and Qimonda (https://www.design-reuse.com/news/15022/ovonyx-qimonda-sign-technology-licensing-agreement-phase-change-memory.html), and others. All of those license obligations were absorbed by Micron when acquiring Ovonyx (https://seekingalpha.com/article/3774746-micron-tainted-love). ECD is still in bankruptcy (http://www.kccllc.net/ecd/document/list/3153).

So, years of R&D and JVs are behind the XPoint Optane(TM) SSDs. They are cross-bar architecture ReRAM arrays of PCM materials, and had the term not been ruined by 17-years of over-promising and under-delivering they would likely have been called OUM chips. Many others tried and failed, but Intel/Micron finally figured out how to make commercial gigabit-scale cross-bar NVMs using OUM arrays. Now they just have to yield the profits…

—E.K.

3D XPoint uses PCM Material in ReRAM Device

IM Flash pre-announced “3D XPoint”(TM) memory for release later this year, and lack of details has led to widespread confusion regarding what it is. EETimes has reported that, “Chalcogenide material and an Ovonyx switch are magic parts of this technology with the original work starting back in the 1960’s,” said Guy Blalock, co-CEO of IM Flash at the 2016 Industry Strategy Symposium hosted by the SEMI trade group. However, contradicting industry terminology conventions, in another article EETimes reported that a spokesperson for Intel has said that, “3D XPoint should not be described as ReRAM.”
First promoted by the master of materials solutions-looking-for-problems Sanford Ovshinsky under the name “Ovonic” trademark, chalcogenide materials form glassy structures with meta-stable properties. With proper application of heat and electrical current, chalcogenides can be made to switch between low-resistivity crystalline and high-resistivity amorphous phases to create Phase-Change Memory (PCM) arrays in silicon circuit architectures. Chalcogenides can also function as the matrix for the diffusion of silver ions in a cross-point device architecture to create a digital “Resistive RAM” (or “ReRAM” or “RRAM”), or create an analog memristor for neuromorphic applications as explored by Prof. Kris Campbell of Boise State in collaboration with Knowm.

Hitachi and Renesas Technology developed Phase-Change Memory (PCM) cell technology employing Ta2O5 interfacial layer to enable low-power operation. (Source: Hitachi)

Hitachi and Renesas Technology developed Phase-Change Memory (PCM) cell technology employing Ta2O5 interfacial layer to enable low-power operation. (Source: Hitachi)

The Figure shows a schematic cross-section of a typical PCM cell. From a scientific perspective, we could say that any memory cell that relies upon a change in material phase to encode digital data should be termed a PCM. However, due to the history of this specific type of PCM device being the only architecture explored for decades (and commercialized for limited niche sub-markets), and due to the fundamentally different circuit architectures, it is reasonable to categorically deny that any cross-point device is a “PCM.”
However, any cross-point memory device based on a resistance change has to be a ReRAM regardless of the switching phenomenon:  phase-change, filament-growth, ion-diffusion, etc. So we could say that this new chip uses PCM material in a ReRAM device.
—E.K.

Cross-point ReRAM Integration Claimed by Intel/Micron

The Intel/Micron joint-venture now claims to have successfully integrated a Resistive-RAM (ReRAM) made with an unannounced material in a cross-point architecture, switching using an undisclosed mechanism. Pilot production wafers are supposed to be moving through the Lehi fab, and samples to customers are promised by end of this year.
HP Labs announced great results in 2010 on prototype ReRAM using titania without the need for a forming step, and then licensed the technology to Hynix with plans to bring a cross-point ReRAM to market by 2013. SanDisk/Toshiba have been working on ReRAM as an eventual replacement for NAND Flash for many years, with though a bi-layer 32Gb cross-point ReRAM was shown at ISSCC in 2013 they have so far not announced production.
Let us hope that the folks in Lehi have succeeded where HP/Hynix and SanDisk/Toshiba among others have so far failed in bringing a cross-point ReRAM to market…so this may be a “breakthrough” but it’s by no means “revolutionary.” Until the Intel/Micron legal teams decide that they can disclose what material is changing resistance and by what mechanism (including whether an electrical “forming” step is needed), the best we can do is speculate as to even how much of a breakthrough this represents.
—E.K.

Moore’s Law is Dead - (Part 3) Where?

…we reach the atomic limits of device scaling.

At ~4nm pitch we run out of room “at the bottom,” after patterning costs explode at 45nm pitch.

Lead bongo player of physics Richard Feynman famously said, “There’s plenty of room at the bottom,” and in 1959 when the IC was invented a semiconductor device was composed of billions of atoms so it seemed that it would always be so. Today, however, we can see the atomic limits of miniaturization on the horizon, and we can start to imagine the smallest possible functioning electronic device.

Today’s leading edge ICs are made using “22nm node” fab technology where the smallest lithographically defined structure—likely a transistor gate—is just 22nm across. However, the pitch between such transistors is ~120nm, because we are already dealing with the resolution limits of lithography using water-immersion 193nm with off-axis-illumination through phase-shift masks. Even if a “next-generation” lithography (NGL) technology were proven cost-effective in manufacturing— perhaps EbDW for guidelines combined with DSA for feature fill and EUV for trim—we still must control individual atoms.

We may have confidence in shrinking to 62nm pitch for a 4x increase in density. We may even be optimistic that we can shrink further to a 41nm pitch for a ~10x increase in density…but that’s nearing the atomic limits of variability. There are many hypothesized nanoscale devices which could succeed silicon CMOS in IC, but one commonality of all devices is that they will have to be electrically connected. Therefore, we can simplify our consideration of the atomic limits of device scaling by focusing on the smallest possible interconnect.

4nmPitchDevice_TheorySo what is the smallest possible electrical interconnect? So far it would be a Single-Walled Carbon NanoTube (SWCNT) doped with metals to be conducting. The minimum diameter of a SWCNT happens to be 0.4nm, but that was found inside another CNT and the minimum repeatable diameter for a stand-alone SWCNT is ~1nm. So if we need three contacts to a device then the smallest device we can build with atoms would be a 3nm diameter quantum dot. As shown in the figure at right, if we examine a plan-view of such a device we can just fit three 1nm diameter contacts within the area.

Our magical device will have to be electrically isolated and so some manner of dielectric will be needed with some minimal number of atoms. Atomic Layer Deposition (ALD) of alumina has been proven in very tight geometries, and 3 atomic layers of alumina takes up ~1nm so we can assume that spacing between devices. A rectangular array would then result in ~16nm2 as the smallest possible 3-terminal device that can be built on the surface of planet Earth.

Note that a SWCNT of ~1 nm diameter theoretically could carry ~25 microAmps across an estimated 5kOhm internal resistance [(ECS Transactions, 3 (2) 441-448 (2006)]. I will leave it to someone with a stronger device physics background to comment as to the suitability of such contacts for useful circuitry. However, from a manufacturing perspective, to ensure electrical contacts to billions of nanoscale devices we generally use redundant structures, and doubling the number of SWCNT contacts to a 3-terminal device would call for ~8 nm pitch.

However, before we reach the 4-8nm pitch theoretical limits of device scaling, we will reach relative economic limits of scaling just one device feature such as a transistor gate. Recall that there are just 22 silicon atoms (assuming silicon crystal lattice spacing of ~0.3nm) across a ~7nm line, and every atom counts in controlling device parameters. Imec’s Aaron Thean recently provided an excellent overview of scaled finFET technologies, and though the work does not look at packing density we can draw some general trends. If we assume 41nm pitch and double fins with 20nm gate length then each device would use ~1,600 nm2.

Where are we now? Let us consider traditional 6-transistor (6T) SRAM cells built using “22nm node” logic process flows to have minimal area of ~100,000 nm2 or ~16,000 nm2 per transistor. At IEDM2013 (9.1), TSMC announced a “16nm node” 6T SRAM with ~70,000 nm2 area or ~10,000 nm2 per transistor.

IBM recently announced that 6 parallel 30nm long SWCNT spaced 8nm apart will be developed as transistors for ICs by the year 2020. Such an array would use up ~1440 nm2 of area. Again, this is at best another 10x in density compared to today’s “22nm node” ICs.

Imec held another Technology Forum at SEMICON/West this year, in which Wilfried Vandervorst presented an overview of innovations in metrology needed to continue shrinking device dimensions. His work with Scanning Spreading Resistance Microscopy (SSRM) is extraordinary, showing ability to resolve 1-2nm conductivity variations in memory cell material. Working with Resistive RAM (ReRAM) material using a 2nm diameter probe tip as the top contact, researchers were able to show switching of the material only underneath the contact…thus proving that a stable ReRAM cell can be made with that diameter. If we use cross-bar architectures of that material we’d be at a 4nm pitch for memory, coincidentally the same pitch needed for the densest array of 3-terminal logic components.

IC SCALING LIMITATION

Pitch / “Node”

Transistor nm2

Scale from 22nm

193nm lithography double-patterning

124nm / “22nm”

16000

1

Atomic variability (economics)

41nm / “7nm”

1600

10

Perfect atoms (physics)

4nm

16

1000

The refreshing aspect of this interconnect analysis is that it just doesn’t matter what magical switch you imagine replacing CMOS. No matter whether you imagine quantum-dots or molecular memories as circuit elements, you have to somehow connect them together.

Note also that moving to 3D IC designs does not fundamentally change the economic limits of scaling, nor does it alter the interconnect challenge. 3D ICs will certainly allow for greater number of devices to be packed into a given volume, so mobile applications will likely continue to pull for 3D integration. However, the cost/transistor is limited by 2D process technologies that have evolved over 60 years to provide maximum efficiency. Stacking IC layers will allow for faster and smaller devices, though generally only with greater costs.

Atoms don’t scale.

Past posts in the blog series:

Moore’s Law is Dead - (Part 1) What defines the end, and

Moore’s Law is Dead - (Part 2) When we reach economic limits.

The final post in this blog series (but not the blog) will discuss:

Moore’s Law is Dead - (Part 4) Why we say long live “Moore’s Law”!

E.K.